Approximating optimal decoding of quantum LDPC codes with narrow frontiers

Abstract: We introduce the Frontier decoder, a pruned dynamic-programming decoder for sparse quantum decoding problems. Frontier processes error variables in a chosen order, merges prefixes with the same residual syndrome and logical label, and approximates logical-coset posterior masses by retaining only a narrow scored frontier. Without pruning, the recursion is exact ordered inference with exponential complexity. In the code-capacity setting, the decoder reaches thresholds close to optimal for the surface code and the color code. In the circuit-level noise model, it achieves state-of-the-art performance with a very small average retained list size: less than 100 for the gross code $[[144,12,12]]$ at a physical error rate of $0.001$. When the list size is constant, the decoder has linear complexity, suggesting the possibility of low-latency implementations.

• The paper presents the Frontier decoder as a pruned dynamic programming method that approximates optimal decoding for quantum LDPC codes.
• It employs merging of boundary states to compress degeneracy and manage candidate error states, achieving near-optimal thresholds under various noise models.
• Empirical results demonstrate state-of-the-art performance with low retained list sizes and transition evaluations, highlighting scalability for fault-tolerant quantum computation.

Frontier Decoder for Quantum LDPC Codes: Technical Overview and Implications

Introduction

This paper introduces the Frontier decoder, a pruned dynamic-programming approach tailored for sparse quantum decoding problems in quantum LDPC codes. The methodology aims to approximate optimal logical maximum-likelihood decoding by efficiently managing the explosion of candidate error states, particularly through exploiting narrow frontiers and boundary degeneracy. The performance is benchmarked against highly specialized decoders, demonstrating both near-optimal threshold behavior in code-capacity regimes and state-of-the-art performance in circuit-level noise models with minimal retained list sizes.

Ordered Dynamic Programming for Quantum Decoding

Quantum LDPC codes present decoding challenges characterized by sparse, sometimes non-local Tanner graphs. Conventional decoders leverage geometric structure (e.g., surface codes [3]), but general LDPC codes require geometry-agnostic techniques. The Frontier decoder formalizes ordered inference: variables are processed according to a chosen order, and prefixes that share identical residual syndromes and logical labels are merged, collapsing degeneracy and reducing redundant computation.

The recursion underlying Frontier, when unpruned, provides exact logical class posterior masses but exhibits exponential complexity. The pruning scheme—retaining only a scored subset of survivors within a specified score gap $\Delta$ and subject to a hard cap $K$—enables practical decoding with typically linear complexity and very low retained list sizes.

Frontier Algorithm and Pruning Strategies

Frontier operates over ordered variables, branching over local error possibilities and merging states sharing boundary labels. Each survivor is assigned a score $S = P + \alpha C$, with $P$ as the log-mass of merged prefixes and $C$ as a suffix-compatibility heuristic score estimating the likelihood that remaining variables can complete the required syndrome. Pruning is governed by the gap $\Delta$ and cap $K$, tuning the tradeoff between accuracy and complexity.

The decoder's geometry-agnostic nature allows direct application to arbitrary parity-check or DEM matrices. After all variables, survivors are aggregated by logical class, outputting the class with maximal retained posterior mass.

Empirical Results and Benchmarking

Performance is evaluated across several families:

• Code-Capacity Regimes: Frontier approximates threshold behavior for rotated surface and color codes under depolarizing noise. For the surface code, thresholds achieve proximity to known optimum ($p \approx 0.189$), matching tensor-network decoders' complexity via frontier size comparisons.
• Circuit-Level Noise: Frontier delivers competitive FERs for both surface codes and BB codes. For gross BB codes [[144,12,12]], at $p=0.001$, the average list size is less than 100, with the median number of transition evaluations in decoding under $6\times10^6$ ($K$0th percentile at $K$1) indicating excellent low-latency profiles.
• Failure Decomposition: Pruning aggressively removes configurations not matching observed syndromes, and increasing $K$2 reduces pruning-induced support loss, shifting failure dominance to terminal ranking (incorrect output among retained logical classes).

Committee variants (forward/backward deadline ordering) further enhance performance, and threshold behavior is geometry-independent across code families.

Relationships to Prior Algorithms

Frontier connects technically to trellis and variable elimination decoders, tensor-network contractions, and sparse search-based decoders [33, 36, 24]. The merging of boundary states before pruning distinguishes it from most search-based algorithms, efficiently encoding degeneracy as a form of compression. The empirical width of Frontier matches bond-dimension cutoffs in tensor-network approaches, but with algebraic boundary definitions rather than geometric ones.

Practical and Theoretical Implications

Frontier enables memory-efficient decoding of quantum LDPC codes, supporting both offline benchmarking and potential real-time implementations due to linear scaling in the number of transition evaluations. The distinction between boundary state compression and explicit representative error lists is practically consequential, particularly for large-scale codes and hardware calibration scenarios.

Terminal aggregation and ranking remain bottlenecks; further enhancement via improved suffix-compatibility scoring or secondary ranking modules could drive accuracy closer to full optimality. The open question is under what structural or noise-model conditions a narrow ordered frontier suffices for capturing logical coset posteriors.

Theoretical implications extend to understanding low-width posteriors in the presence of quantum degeneracy and to the optimality of boundary state merging in general factor graph models.

Future Directions

Key challenges include: optimizing input ordering, exploring committee extensions, integrating detector correlations into suffix scoring, and generalizing to logical circuit models with non-Pauli errors. The practical applicability for online decoding depends on aligning ordering with circuit temporal structure.

Further, deep understanding of when small boundary width yields accurate posteriors warrants investigation, analogous to bond-dimension analyses in tensor-network decoding.

Conclusion

The Frontier decoder provides a principled, adaptable approach to quantum LDPC decoding, leveraging ordered dynamic programming and efficient degeneracy compression. Empirical benchmarking demonstrates threshold proximity with minimal retained list size, validating both code-capacity and circuit-level applications. The technique offers efficient logical inference without geometry dependency, and its low-latency attributes suggest viability for scalable fault-tolerant quantum computation. The interplay between pruning strategies, degeneracy collapse, and terminal aggregation defines an open space for future improvements and theoretical exploration.